Modern organizations are universally drawn to the allure of artificial intelligence (AI) and machine learning (ML). We live in a world where vast amounts of data are being collected and unprecedented compute power is available to extract value. The advancement of technology in machine learning and data science can truly transform industries through insights and predictions.
But like anything else, the great power of AI requires great responsibility. Without taking a diligent approach, many AI projects and initiatives will fail to create business value. Creating an AI strategy is a smart way to make sure your early projects deliver business value and lead to continued growth and investment.
In this guide, we’ll dive into why many AI strategies fail, explore the benefits of building a proper AI strategy, and finally, offer a step-by-step guide to help your business build a successful AI strategy.
AI and ML initiatives without a strategy have a tendency to fail, but they don’t always fail in the same way. Some failures are sudden and spectacular, but the most costly failures are those that take a long time to unfold. Here are some common failure modes that we’ve seen.
Any AI project will require investment. To justify that investment, it is important to estimate the potential business value of the solution before making significant development investments. We’ve seen teams get ahead of themselves and excitedly develop solutions, but never estimate how valuable those solutions will be.
Many ML projects are spawned based on external inspiration. For example, someone might have seen a cool image-processing application at a trade show and aim to replicate that within their organization. But the novelty of a solution doesn’t necessarily mean it aligns with the problems facing your organization. Before developing a solution with ML, try to imagine how you would solve that problem without ML. Is this still an interesting problem for your organization to solve?
Operationalizing AI requires solutions to be deployed into a production-grade environment. In other words, AI projects are really software projects, but with extra complexity. We’ve often seen solutions developed by data scientists, but without the infrastructure or organizational support to take their solution to production. If you’re developing an ML solution to solve a problem with AI, make sure you have the teams in place to carry that solution forward.
AI projects often require significant research and development effort. Research is only valuable if it is reproducible. Are your data scientists working in siloed environments (worst case: their laptops) and not versioning their code and results in a central location? If so, there is a great risk of their effort going to waste. If those data scientists move into different roles, that work is instantly lost. When it comes time to deploy those models to production, reproducibility is key for transferring solutions to engineering and operations teams. Make sure to identify best practices and provide infrastructure to ensure data scientists are working in a reproducible manner. Using a Model Registry is the first step!
Sometimes ML solutions succeed in becoming operationalized, but fail to obtain production-grade status due to a lack of automation. Models always need to be updated and redeployed. If teams are manually retraining models and deploying artifacts, these solutions will eventually become too cumbersome to maintain. Make sure to include DevOps automation principles in your MLOps stack.
An ML solution is only providing value to your business if you can clearly measure that value. Without monitoring and observability, you won’t be able to report on the success of your solutions and justify your investments. Additionally, you’ll lack the information necessary to improve the solution or resolve issues if they arise. Make sure to build your solution in a way that enables monitoring, alerting, reporting, and evaluation.
Having read about all the ways projects can fail, it might be clear why a strategic approach is essential. Rather than relying on negative counterexamples, here are a few reasons why it is helpful to to build a strategy for your AI projects.
First off, you’ll be able to prioritize your potential projects based on the relative effort and estimated ROI. In doing so, you’ll make sure your first (or next) project has potential to deliver a clear and quick win for your organization. This first project can then become a success story that helps to justify more investment in AI and ML initiatives.
Second, you can make decisions that will prevent the accumulation of technical debt. An AI strategy includes architectural and best-practice guidance that will help data scientists and machine learning engineers develop robust solutions. In particular, your strategy should select best-of-breed technologies that are ideally suited to your organizational strengths and weaknesses. If you don’t set an architectural roadmap, decisions made on early projects are more likely to come back to haunt you.
Finally, creating a strategy up front will set the stage for quickly operationalizing AI solutions once they’ve been developed. As a result, you’ll be able to deliver business value as quickly as your data science teams can innovate. There’s nothing worse than seeing clear potential in a solution, but then waiting months or years to capitalize on it.
The first step in building an AI strategy that delivers value is discovery. In this phase, you’ll collect information about your organization that will guide strategic decisions. Don’t be fooled into thinking that discovery is boring – it can actually be the most exciting and invigorating phase! Due diligence in discovery will also set a strong foundation for everything to follow. Discovery efforts fall into two categories: organizational discovery and use-case discovery.
Organizational discovery uncovers the details that make your organization unique. What are the most important business priorities for the upcoming year and beyond? Knowing business priorities will help select use cases. How large is your IT department and how deep are your technical capabilities? Selecting technologies will depend on this information. And how will organizations develop and support AI solutions? It is essential to understand whether solutions will be owned by central IT or within business units.
Use-case discovery is where the discovery process really gets exciting because you’ll actually be laying down ideas for potential AI solutions. Much of the information will be gathered through interviews with business units. In these interviews, you should identify the most important problems and pain points that can be solved through AI. Data scientists can then identify solutions that can be solved with AI.
For each use case, there are important questions that must be asked to determine viability of the solution. Make sure to understand the following about each use case:
AI solutions don’t just spring out of the lab and into production. Operationalizing solutions requires some key infrastructure components to support data science and ML engineering teams. It is important to make sure you have an architecture that makes sense for your organization.
Having a clear reference architecture will assist in selecting technologies in the next phase. It will also make sure you don’t miss any steps in the MLOps cycle that would prevent you from creating complete solutions. For a complete overview of MLOps, make sure to check out our comprehensive guide or beginner’s introduction.
A reference architecture consists of multiple diagrams and documents that provide different views of your system. Having multiple views helps audiences from various backgrounds understand the architecture – sort of like zooming in and out on a map. We recommend that a reference architecture include at least the following elements:
Capabilities – Your architecture should include a diagram or document that describes the capabilities of the system. Rather than focusing on specific technology choices, this view helps describe the value provided by the platform. The capabilities included will depend on the use cases and skill level that exists within your organization. Clearly demonstrating capabilities will help obtain buy-in for your platform.
Technologies – Technologies are key to enabling capabilities. Your technology choices should be clearly documented with written explanations, as well as alternatives that were considered but rejected.
Development Practices – Common development practices are key to agility and scale. To help your organization deliver repeatable success, you should outline DevOps and MLOps practices for consistent pipelines and automation. For AI and ML, this means creating guardrails on how models are developed, tracked, and promoted to production.
Compliance and Governance – Different industries have different requirements for AI and ML. In some cases, such as HIPPA in health care, data is highly regulated and specific requirements must be met for use. In other domains, ML predictions themselves have special regulations, such as a right to explanation for any automated decision-making in credit reporting.
Your reference architecture should be very complete, and to an extent, it should be aspirational. Your first projects and use-cases may not use every element of the reference architecture. In these early initiatives, you may be able to bypass certain elements to deliver a solution quickly. But ultimately, having a reference architecture will make sure that you’re able to complete the relay even while you step around a few hurdles.
In terms of AI capabilities and technologies, you’ll want to think about a few key components:
You’ll also want to relate these components to your organizational strengths and weaknesses. Which of these components does it make sense to develop and support in-house? Which would be better to purchase or have supported as SaaS?
The AI and ML space is increasingly crowded and highly segmented. There are exciting and wonderful products coming on the market every day, but not every one of these is right for your organization. You’ll want to make sure to identify vendors that truly compliment your organization’s strengths and weaknesses.
Your reference architecture will be a guiding light in this phase. For each element in your reference architecture, you’ll want to create a list of potential tools and vendors. When assessing each, ask yourself:
You should also prioritize the procurement of these tools. Which tools will be necessary for the early phases of your projects? Which ones will be necessary later? Procuring and integrating tools takes time and effort, so you’ll want to make sure you build out your architecture in an orderly fashion.
No organization is completely prepared to pursue its first AI initiatives. Your organization will likely have gaps in certain domains and skills. For instance, some organizations have great software and IT departments but are weak in terms of statistical modeling and machine learning. Other organizations are great from a scientific and R&D perspective but lack the engineering skills or operational expertise to fully deploy solutions.
Ask yourself whether it makes sense to develop all of these skills internally. Hiring and training can take a long time. Developing new competencies can also be a distraction from the things you do best. If you are going to develop new competencies, is your leadership fully onboard with the level of investment?
Alternatively, you may wish to outsource certain steps in the process. Perhaps it makes sense to tackle R&D within your organization to own your intellectual property but outsource the deployment and operations to an external firm. Organizations with strong engineering and IT teams could of course seek to do exactly the opposite!
Many AI solutions are intended to automate inefficient tasks or job functions within your organization. You’ll of course want to factor in the related cost savings into your planning. Take care, however, to be sensitive about the potential organizational and personnel impacts of your solution. Will positions be eliminated? How many departments will be affected?
The previous steps have gathered all relevant information for your AI initiatives; now it’s time to build a roadmap. You will want to build a roadmap that prioritizes quick wins to demonstrate business value and justify investments, both current and future.
There will be costs associated with each phase in your roadmap. Calculate these costs carefully and clearly document when they will be incurred. Presenting costs transparently and honestly will be vital to continued investment in AI initiatives.
Build your roadmap by making the following prioritization decisions:
An example AI Development Roadmap
Making the choices above should help your roadmap start to come into focus. Make sure to document which teams are involved at each phase in the roadmap and clearly state their roles and responsibilities. The roadmap will ultimately serve as a communication tool, so you’ll want to make sure the relevant stakeholders are identified.
Roadmaps are important, but that doesn’t mean everything is fixed in time and place. Even with a clear strategy, your projects will depend on questions that you haven’t answered yet. Some of these are research questions. Make sure that you build your roadmap with appropriate feedback mechanisms and be clear about moments when you may have to pivot.
AI projects usually follow an iterative lifecycle. Data Science teams will start with data discovery for a particular use case, where they gather and explore available data. That data is then used to train ML models, optimize them, and validate those models as a solution for the use case at hand. If validation is successful, those models can be deployed and monitored as AI solutions. As projects mature, data scientists move back to the discovery phase to identify new solutions or areas for improvement on existing ones.
The Agile AI lifecycle
Creating an AI strategy is a very rewarding process, but of course, it’s just the beginning. Presenting your plan to leadership will make sure that everyone is aligned with your strategy, and also help prevent you from taking wrong turns and wasting investment. This will be an exciting time to celebrate your undertaking and imagine a future empowered by AI!
We hope this guide helps you in advancing AI initiatives at your organization. You will likely have to tailor some steps based on your experience and unique goals, but the overall framework should apply to most businesses..
At phData, we believe that AI is important for any modern organization, and we’re here to help. We’ve got years of experience in data science, MLOps, and data engineering. If you’d like to leverage our insights in your AI initiatives, we’d love to get started.